NOT FOR QUOTATION WITHOUT PERMISSION OF THE AUTHOR
DIRECTIONAL DIFFERENTIABILITY OF A CONTINUAL MAXIMUM FUNCTION OF QUASIDIFFERENTIABLE FUNCTIONS V . F . Demyanov
I.S. Z a b r o d i n J u n e 1 9 8 3 WP-83-58
V o r k i n q P z p e r e
a r e
i n t e r i m r e p o r t s on work o f the I n t e r n a t i o n a l I n s t i t u t e f o r A p p l i e d Systems A n a l y s i s and have r e c e i v e d o n l y l i m i t e d r e v i e w . Views o r o p i n i o n s e x p r e s s e d h e r e i n do n o t n e c e s s a r i l y r e p r e - s e n t t h o s e o f the I n s t i t u t e o r o f i t s N a t i o n a l Member O r g a n i z a t i o n s .INTERNATIONAL IXSTITUTE FOR APPLIED SYSTEMS ANALYSIS A-2361 L a x e n b u r g , A u s t r i a
PREFACE
Much recent work in optimization theory has been con- cerned with the problems caused by nondifferentiability.
Some of these problems have now been at least partially overcome by the definition of a new class of nondifferen- tiable functions called quasidifferentiable functions, and the extension of classical differential calculus to deal wit11 this class of functions. This has led to increased theoretical research in the properties of quasidifferen- tiable functions and their behavior under different con- di tions .
In this paper, the problem of the directional differen- tiability of a maximum function over a continual set of
quasidifferentiable functions is discussed. It is shown that, in general, the operation of taking the "continual"
maximum (or minimum) leads to a function which is itself not necessarily quasidifferentiable.
Andrze
jWierzbicki Chairman
System
&Decision Sciences
DIRECTIONAL DIFFERENTIABILITY OF A CONTINUAL MAXIMUM FUNCTION OF QUASIDIFFERENTIABLE FUNCTIONS V. F . Demyanov
I.S. Zabrodin
1. INTRODUCTION
Optimization problems involving nondifferentiable functions are reccgnized to be of great theoretical and practical sig- nificance. There are many ways of approaching the problems caused by nondifferentiabi.lity, Some of which are now quite well developed. while others still require much further work. A
comprehensive bibliography of publications concerned with non- differentiable optimization has recently been compiled [1]--
major contributors in this field include J.P. Aubin, F.H. Clarke, Yu.M. Ermoliev, J.B. Hiriart-Urruty, A.Ya. Kruger, S.S. Kutateladze, C. ~emargchal, B.
S. Morduchovich, E .A. Nurminski, B. N. P~her~ichniy, R.T. Rockafellar, and J. Warga.
The notion of subgradient has been generalized to nonconvex functions in a number of different ways. One of these involves the definition of a new class of nondifferentiable functions
(quasidifferentiable functions) which has been shown to represent a linear space closed with respect to all algebraic operations as well as to the taking of pointwise maximum and minimum
[ 2 , 3 ].
This has led to the development of quasidifferential calculus--
a generalization of classical differential calculus--which may
be used to solve many new optimization problems involving non-
differentiability
[ 41 .
This paper d e a l s
w i t ht h e problem of t h e d i r e c t i o n a l d i f - f e r e n t i a b i l i t y of a
maximumf u n c t i o n o v e r a c o n t i n u a l
s e tof q u a s i d i f f e r e n t i a b l e f u n c t i o n s .
I t w i l l b eshown t h a t i n g e n e r a l t h e o p e r a t i o n of t a k i n g t h e " c o n t i n u a l "
maximum (minimum) l e a d st o a f u n c t i o n which
i si t s e l f n o t n e c e s s a r i l y q u a s i d i f f e r e n t i a b l e .
2 . AUXILIARY RESULTSL e t
us c o n s i d e r a mapping
G : En-
2Em , where
2Em
d e n o t e s t h e s e t of a l l s u b s e t s of
,E. Fix x
0 E Enand
9
E En, Ogm=l. Choose
y E G( x
)and i n t r o d u c e t h e s e t
0 1
W e
s h a l l d e n o t e t h e c l o s u r e
o f y( y ) by
F( y ) ,
i.e . ,
The set
F( y )
isc a l l e d t h e s e t of f i r s t - o r d e r f e a s i b l e d i r e c t i o n s a t
thep o i n t y
EG(x
)i n t h e d i r e c t i o n g .
0
R e m a r k 1.
I n
thec a s e where
Gdoes n o t depend on x , t h e s e t
F( y )
= F (x, g , y ) does n o t depend on x and g , and
i s0 0
a
cone c a l l e d
thecone of f e a s i b l e d i r e c t i o n s a t y .
A
mapping
G i ss a i d t o a l l o w f i r s t - o r d e r approximation a t a p o i n t x i n
thed i r e c t i o n g
E En, YgI-1 , i f , f o r an
0
a r b i t r a r y convergent sequence { y k } such t h a t
t h e f o l l o w i n g r e p r e s e n t a t i o n h o l d s
:where vk
EF ( x , g , y )
0 IakVk -
0 I !? 5 ( x 0 )I n what f o l l o w s
i t i sassumed t h a t t h e mapping
G iscontinuous ( i n t h e Hausdorff m e t r i c ) a t a p o i n t x and
0
allows f i r s t - o r d e r approximation a t x i n any d i r e c t i o n
0
g E E n , Ilgll
= 1.
I t i s
a l s o assumed t h a t f o r every
t h e s e t G(x)
i sc l o s e d and s e t s G(x) a r e j o i n t l y bounded on
S ( X ), i . e . , t h e r e e x i s t s an open bounded
s e t B C Em0
such t h a t
L e t
us c o n s i d e r t h e f u n c t i o n f ( x )
=max $ ( x , y )
YfG
( x )
where f u n c t i o n
@( z )
=4 ( x , y ) i s continuous i n z
= [x , y ] on S 6 ( x 0 )
x Band d i f f e r e n t i a b l e on
Zi n any d i r e c t i o n
0
rl =
[ g , q ]
E En+,, i . e . , t h e r e e x i s t s a f i n i t e l i m i t
Here
Suppose t h a t t h e f o l l o w i n g c o n d i t i o n s h o l d :
C.-t;.dition
1.I f
qk-
qt h e n
C o n d i t i o n 2 . L e t yk E R ( x
+
a k g ),
yk-
y-
s i n c e0
G a l l o w s f i r s t - o r d e r a p p r o x i m a t i o n a t x .
,
t h e n0
I t i s assumed t h a t the q k l s
are
bounded.C o n d i t i o n 3. F u n c t i o n 4 i s L i p s c n i t z i a n i n some neigh- borhood of t h e s e t Z
.
0
Then t h e f o l l o w i n g r e s u l t h o l d s .
T h e o r e m I . T h e f u n c t i o n f is d i f f e r e n t i a b l e at the point x in the d i r e c t i o n g a n d
0
a f ( x
a @ ( x
, Y )O = sup SUP 0
ag ~ E R ( X ) O q f r ( y ) a [ g , q l
Proof. L e t us d e n o t e by A t h e r i g h t - h a n d s i d e of (1). F i x y E R ( x ) and q E y ( y )
.
Then y+
a q E G ( x + a g ) f o r0 0
s u f f i c i e n t l y s m a l l a > 0 and
a o
( X , Y ) f ( X+
a g )a
@ ( x +ag , y + a q ) = f ( x + a 00 0 0
Hence
1
a o
(X , y )l i m
h ( a )l i m
- - -
[ f ( x+
a g )-
f ( x1 1
0a-tO a+O a o o
a t MI
S i n c e y E R ( x ) and q E y ( y ) a r e a r b i t r a r y t h e n
0
ao
c x 0 , y )-
liln 1 ( a ) 2 supa-+O SUP
;.fRix ) qEy ( y ) a [ q , q l
0
L e t qk
-
q,
qk E y ( y ).
Then q E r ( y ).
I t f o l l o w s from C o n d i t i o n 1 t h a tBut
a 4 ( x
, Y )a 4 ( x
, Y )-
a[ gt4*1<
4 E y s u p ( Y ) a [ g t s 0I
S i n c e
r
( y ) = c l Y ( y ),
then from ( 3 )a @ ( x 0 , y ) a $ ( x 0 t y )
a o ( x o
/ y )SUP 4 s u p s u p
* ( Y ) a [ g , q I W ( Y ) a ~ g ~ q l
GY
( Y )a [
g , q lFrom
a q ( x , Y )
a w
, Y )suy:
A
= s u p 0+ r ( y ) a [ g l q i +Y ( Y Y a [ g , q l
and f 011 ows that
NOW l e t u s c h o o s e s e q u e n c e s { y k } and { a k } s u c h t h a t
The c o n d i t i o n s imposed on t h e mapping G and t h e con- t i n u i t y of t h e f u n c t i o n e n s u r e t h a t the f u n c t i o n
f
i sc o n t i n u o u s a t x o x e n c e , from t h e e q u a l i t y f ( x +'ikg) = o ( x t C l , g
,
pki ,0 0
- -
o n e c a n c o n c l u d e t h a t f ( x ) = ~ ( x , y )
,
e y E R ( x ).
0 0 0
Since the mapping G allows first-order approximation at x , then yk
=- y
+akqk + o (ak) , where. qk
0
E
r(?) ,
akqk
---40. From Conditions 2 and 3 the qk' s are bounded and
1.;:function is Lipschitzian around Z . Without loss of
0
generality one can assume that qk - q . It is clear that q
Er ( y ) . Hence
where
Since
@is a Lipschitzian function, then
It follows from
(6)-(8)that
.front wnicn it is clear that
a$(x ,Y)
T r n h(a)
SUPSUP
o= A .
Comparison of
( 5 )and (9) now shows that lim h(a) exist-
a+c!ana is equal to A , thus completing the proof.
Remark 2.
Equation (1) has been proved under some different
assumptions elsewhere
[ 51 (see also
[ 61 ,
$10
). The case where
4is differentiable was studied by Hogan
[ 7 ] .3. QUASIDIFFERENTIMLE CASE
L e t us c o n s i d e r once a g a i n t h e f u n c t i o n f ( x = rnax Q ( x , y )
a yEG
( X I
where mapping G s a t i s f i e s t h e c o n d i t i o n s s p e c i f i e d e a r l i e r and f u n c t i o n Q ( 2 ) = Q ( x , y ) i s continuous i n z on S ( x B
g 0
and q u a s i d i f f e r e n t i a b l e on 2. , i . e . , f o r a n y p a i n t z = [ x
0
-
0o f Y o l
E 0t h e r e e x i s t convex compacts
- 20
( 2 a ) C En+, anda + ( ~
ac
E ~ + ~such t h a t
+
n i n [ (wltg) + ( w 2 f g )I [ W ~ ' W ~ I ~ ~ ~ ( ~
0 )I t
i s
a l s o assumed t h a t C o n d i t i o n s 2 and 3 a r e s a t i s f i e d . (Condition 1 f o l l o w s immediately from ( l l ) . ) Thus, a l l t h e c o n d i t i o n s of Theorem 1 a r e f u l f i l l e d and we a r r i v e a tT h e o r e m 2. T h e f u n c t i o n f d e f i n e d b y ( 1 0 1 i s d i r e c t i o n a l l y d t j f s r e n t i a b l z a n d , m o r e o v e r ,
a f ( ~
1A = sup SUP
\
max [ ( v l I g ) + ( v 2 , q )I
ag Y-(x o )
W ( Y ) 1
[ v l f v 2 1 e 2 Q ( x-
0 , y )+
min [ (wS&) + (w2tq)1 I / .
( 1 2 )[ w1,w21
e a 4
( x , y )0
R e m a r k 3 . S i n c e y E R(x , y )
,
t h e f o l l o w i n g r e l a t i o n h o l d s :0
SUP
\
max ( v 2 ' q ) + n i nI
( w 2 ' q ) j = 0 + r ( y )
1
~ ~ € 3 0 ( X , y ) -- Y 0 w2ES$y (x 0 I Y )
Here 39 (x ,y) and -
d $(xo,y) are the projections of sets
- Y
0- Y
ap (xo ,y) and a$ (xO ,y), respectively, onto Em . -
R e m a r k 4 .
Pshenichniy
181 considered the case where G (x) does not depend on x and F (x)
=4 (x,y) is a directionally differentiable function for every fixed y Y , i.e., there exists
a9 (x,Y) 1
=
lim -
[$(x+ag,y) -
$(x,y)l
ag
a 4 0a
Then
Under an additional assumption about the behavior of o(a,y) in (13) , it has been proved that
af (XI a4 (x,Y)
= max
ag yER (x) ag
It is clear that equation (14) differs from equation (12).
E x a m p l e I .
Let x E E l , y E E 1 ,
G ( x ) G =[-2,2] ,
4(x,y)
=x - 21y - xl , and
f(x)
=max (x - 2 I y - X I ) -
yEl -2,21 It is clear that
Choose x
E(-2,2) and verify equation (14) . We shall now compute the right-hand side of (14). Since
4(x,y)
=x - 2 max {y - x , -y + x) , then for y
E R ( x ) ={XI
it follows
[ 9I that
a$ (x,Y)
=
g - 2 max E-g,g) .
a4
Hence f o r gl = +l
a +
( x , Y )max
= 1 - 2 2 - 1 ,F R ( x )
and f o r g 2 = -1
But from ( 1 6 ) i t
i s
c l e a r t h a tThus e q u a t i o n ( 1 4 ) d o e s n o t h o l d f o r any d i r e c t i o n g ( i n E l t h e r e a r e o n l y two d i r e c t i o n s g s u c h t h a t llgll = 1 : g = +1 and g = -1).
sow l e t u s v e r i f y e q u a t i o n ( 1 2 ) . Denote by D the t i g h t - hand s i d e of ( 1 2 )
.
The f u n c t i o n $ ( x , y ) i s q u a s i d i f f e r e n t i a b l e . From q u a s i d i f f e r e n t i a l c a l c u l u s [ 2 - 4 1 it f o l l o w s t h a t i f y =x
t h e n o n e c a n c h o o s e a $ ( x , y )
-
= { ( 1 , O ),
5 $ ( x , y ) = c o { ( - 2 ~ 2 ) I ( 2 1 - 2 )1 -
For t h e f u n c t i o n f d e s c r i b e d by ( 1 5 ) w e have
Computing D :
I
D
= sup1
( 1 . g )+
( 0 . q )+
inin [(wl.g) + (w2.q)1j
-1 ["l' w 2 lEcoI(-2,2) , ( 2 , 2 )
I
= g
+
sup min [ ( ~ ~ - 9 ) + (w,41.
( 1 8 )$El [ w1
,",I - s[
( - 2 , 2 ),
( 2, I
I t i s c l e a r from F i g u r e 1 t h a t f o r any g t h e second
term
on t h e r i g h t - h a n d s i d e of ( 1 8 ) i s e q u a l t o z e r o , i . e . , D = g.
(The supremum i n ( 1 8 ) is a t t a i n e d a t q = g . )
F i g u r e 1.
T h u s , from ( 1 7 )
,
e q u a t i o n ( 1 2 ) is c o r r e c t ir, t h i s c a s e .Remark 5 . When s o l v i n g p r a c t i c a l problems i n which i t i s r e q u i r e d t o minimize a max f u n c t i o n o v e r a c o n t i n u a l s e t of p o i n t s , t h i s maximum f u n c t i o n i s o f t e n d i s c r e t i z e d ( t h e con- t i n u a l s e t r e p l a c e d by a g r i d of p o i n t s ) . I n many c a s e s t h i s o p e r a t i o n i s a l e g i t i m a t e one
[ l o ] ,
b u t w e s h a l l show t h a t i n the c a s e where i s a q u a s i d i f f e r e n t i a b l e f u n c t i o n t h i s r e p l a c e m e n t may b e d a n g e r o u s .L e t f a g a i n b e d e s c r i b e d by ( 1 5 )
.
D e f i n e f N a swhere
aN
= {x1 I . =+ I
I Xk E [ - 2 , 2 1.
T h i s f u n c t i o n h a s N l o c a l minima (see F i g u r e 2 1 , a l t h o u g h t h e o r i g i n a l f = x h a s no l o c a l minimum which i s n o t a l s o g l o b a l o n [ - 2 , 2 ]
.
T h i s d e m o n s t r a t e s t h a t t h e d i s c r e t i z a t i o n o f t h e max-type f u n c t i o n must b e c a r r i e d o u t v e r y c a u t i o u s l y .F i g u r e 2
Ezample 2 . T h i s example i l l u s t r a t e s t h a t C o n d i t i o n 2 i s e s s e n t i a l t o o u r argument.
9 ( x , Y ) = x
-
2 n i n i / ( x -t 3 1 2
+ ( y-
t ) 2tEE - 2 , 2 1
f ( x ) =
max
@ ( x , Y ) yEI - 2 , 2 1I t i s c l e a r t h a t f o r x E ( - 2 , 2 )
,
I f y =
\3J;- ,
t h e n t h e minimum i n ( 1 9 ) i s a c h i e v e d a t t = x.
Take x = 0
.
Then R ( 0 ) = {O},
I'(y) = E l.
C o n s t r u c t a0
q u a s i d i f f e r e n t i a l of t h e f u n c t i o n $ a t t h e p o i n t ( 0 , O )
.
By t h e r u l e s o f q u a s i d i f f e r e n t i a l c a l c u l u s one c a n choose
L e t u s d e n o t e by D t h e r i g h t - h a n d s i d e of e q u a t i o n ( 1 2 ) and e v a l u a t e it.
D = 3 ( g . ) = s u p min (wl.g + W2.q)
I
+El Lwltw21~co [ ( - 2 t O ) t ( 2 , O ) I
I
If g1 = 1 t h e n D(gl) = -1 if g 2 = -1 t h e n D ( g 2 ) = - 3
.
But i t i s c l e a r from ( 2 0 ) t h a t a f (O)/ag = g
.
Thus e q u a t i o n ( 1 2 ) d o e s n o t h o l d , and t h e r e a s o n i s t h a t C o n d i t i o n 2 i s n o t s a t i s f i e d . I n d e e d , t a k i n g an a r b i t r a r y s e q u e n c e xk = x+
akg0
where ak- + O
,
p u t t i n g , f o r example, g = 1,
x = 0,
w e0
o b t a i n yk =
+
a vk k t Y k E R ( x k )
.
For y
-
= 0 t h i s l e a d s t o R ( x k ) = {V
a k }.
ButREFERENCES
E.A. Nurminski ( E d . ) . P r o g r e s s i n N o n d i f f e r e n t i a b l e O p t i m i z a t i o n . CP-8.2-S8, IIASA, Laxenburg, 1982.
V.F. Demyanov and A.M. Rubinov. On q u a s i d i f f e r e n - t i a b l e f u n c t i o n a l s . S o v i e t Math. Dokl., 2 1 (1)
,
pp. 14-17, 1980.
L.N. Polyakova. Necessary c o n d i t i o n s f o r an extremum of q u a s i d i f f e r e n t i a b l e f u n c t i o n s . V e s t n i k Leningrad U n i v e r s i t y , No. 1 3 , pp. 57-62, 1980.
V.F. Demyanov and L.V. V a s i l i e v . N o n d i f f e r e n t i a b l e O p t i m i z a t i o n . Nauka, Moscow, 1981. ( I n R u s s i a n ) T.K. Vinogradova, V.F. Demyanov, and A.B. Pevnyi.
On the d i r e c t i o n a l d i f f e r e n t i a b i l i t y of f u n c t i o n s of maximum and minimum. O p t i m i z a t i o n ( N o v o s i b i r s k )
,
1 0 ( 2 7 )
,
pp. 17-21, 1973. ( I n R u s s i a n )V.F. Demyanov. Minimax: D i r e c t i o n a l D i f f e r e n - t i a b i l i t y . Leningrad U n i v e r s i t y P r e s s , L e n i n g r a d , 1974. ( I n R u s s i a n )
W . W . Hogan. D i r e c t i o n a l d e r i v a t i v e s f o r extreinal v a l u e f u n c t i o n s . W e s t - Manage. S c i . I n s t . ,
Working Paper No. 1 7 7 , Los Angeles, 1971.
B.N. P s h e n i c h n i y . Necessary C o n d i t i o n s f o r Extremum Problems. Marcel Dekker, New York, 1 9 7 1 .
V.F. Demyanov and V . N . Malozemov. I n t r o d u c t i o n t o Minimax. Wiley, New York, 1974.
V . N . Malozemov. On t h e convergence of a g r i d method i n t h e b e s t polynomial approximation problem. V e s t n i k Leningrad U n i v e r s i t y , No. 1 9 , pp. 138-140, 1970.
( I n R u s s i a n )